Intel Updates Developer Toolkit for HPC
Intel today updated its developer toolkit for high performance computing, a step the company believes will make supercomputing faster and more approachable to non-experts in this era of high demand for expertise in everything from big data and analytics to HPC and computer science.
Intel Parallel Studio XE 2016 focuses on improving performance, leveraging standards, and automating some tasks. The update includes the Intel Data Analytics Acceleration Library; Vectorization Advisor, and MPI Performance Snapshot, plus support for recent standards, operating systems, and processors.
The Data Analytics Acceleration Library (DAAL) includes advanced analytics libraries that support all data analysis stages, connect simply to object-oriented APIs for C++ and Java, and easily connect to analytics platforms such as Hadoop and Spark, said James Reinders, chief evangelist of developer products at Intel, in an interview.
"Intel DAHL makes [HPC] more approachable because data scientists can use this tool," he said. "I think with all of these tools, features, I see a very strong emphasis from us to make tools that are approachable to people who are new to the [HPC] community. It may not be their whole career. They may be asked to work on a big data project or something."
In beta since the beginning of the year, early customers use DAAL to solve computational problems as part of big data analysis, said Reinders.
"One of our customers is trying to correlate the different services a customer chooses and look at what traditional finance products would be best to target at that customer," said Reinders. "To do that they make heavy use of correlation capabilities to be able to process a lot of data and correlate data for consistencies that show buying patterns."
The Vectorization Advisor guides software developers' use of vectorization to maximize the performance of their code and accelerate the software. But attaining vectorization's rewards can be challenging and typically developers either do the process manually or turn to external consultants, Reinders said.
"The tool will allow a software developer to pull up code, do some analysis, and start clearing out opportunities for vectorization, explain which areas are probably more profitable to start vectorization, and start to explain why some areas of code are probably not beginning to vectorize," he said. "The challenge of developing your code and figuring out why it's not vectorizing seems to be a needlessly daunting task. We've tried to distill it all in one place … to increase the efficiency of the programmer."
With MPI Performance Snapshot, Intel provides scalable profiling for MPI and hybrid clusters in a lightweight overhead that profiles to 32,000 Ranks. The tool supports CPI ad memory-bound performance metrics and now includes an application summary in the HTML report, new command line options, and an MPS tool for statistical analysis for Windows.
"MPI Performance Snapshot collects a smaller amount of data – 32000 Ranks – in a manner that feels good, performance-wise. At 32000 Ranks it's a very responsive tool for looking at what's really going on in a program," said Reinders. "This is a very exciting feature for people who are trying to tune at very large scale of machines for very large science or very large big data problems."
To simplify adoption and integration, Intel adhered to multiple existing (and one beta) standards, operating systems, and compilers. Supported standards include Unicode strings and C11 anonymous unions and enhanced C++14 standard support; Windows 7 through 10 and Windows Server 2008-2012; multiple flavors of Linux such as Debian, Fedora, Red Hat, SuSE, and Ubuntu, and Apple OS X. Additionally, version 2016 supports several Fortran 2008 submodules and features from draft Fortran 2015. Finally, the developer tuned the toolkit to take advantage of its processors, such as those use Skylake and Knights Landing microarchitecture and AVX-512, Intel said.
"It's even more exciting when you follow the draft standards. It gives people an opportunity to try them and give people the opportunity to provide feedback to the standards committees as well," said Reinders.
To support the toolkit, Intel will host a series of webinars between Sept. 1 and Nov. 10. The webinars, which will be available on demand, cover topics such as an introduction to Parallel Studio XE 2016, big data analytics and the tool, and tuning multi-level parallelism. The toolkit rollout coincides with publication of High Performance Parallelism Pearls, Volume Two, which includes 73 expert contributors in 24 contributed chapters who share examples of standard parallelism models.